anthropic support - dymanic window size / window popup (#2284)

skyvern/forge/agent.py

@@ -7,7 +7,7 @@ import string
 from asyncio.exceptions import CancelledError
 from datetime import UTC, datetime
 from pathlib import Path
-from typing import Any, Tuple
+from typing import Any, Tuple, cast
 
 import httpx
 import structlog
@@ -72,6 +72,7 @@ from skyvern.forge.sdk.workflow.context_manager import WorkflowRunContext
 from skyvern.forge.sdk.workflow.models.block import ActionBlock, BaseTaskBlock, ValidationBlock
 from skyvern.forge.sdk.workflow.models.workflow import Workflow, WorkflowRun, WorkflowRunStatus
 from skyvern.schemas.runs import CUA_ENGINES, CUA_RUN_TYPES, RunEngine
+from skyvern.utils.image_resizer import Resolution
 from skyvern.utils.prompt_engine import load_prompt_with_elements
 from skyvern.webeye.actions.actions import (
     Action,
@@ -1428,6 +1429,7 @@ class ForgeAgent:
         ]
         thinking = {"type": "enabled", "budget_tokens": 1024}
         betas = ["computer-use-2025-01-24"]
+        window_dimension = cast(Resolution, scraped_page.window_dimension) if scraped_page.window_dimension else None
         if not llm_caller.message_history:
             llm_response = await llm_caller.call(
                 prompt=task.navigation_goal,
@@ -1437,6 +1439,7 @@ class ForgeAgent:
                 raw_response=True,
                 betas=betas,
                 thinking=thinking,
+                window_dimension=window_dimension,
             )
         else:
             llm_response = await llm_caller.call(
@@ -1446,6 +1449,7 @@ class ForgeAgent:
                 raw_response=True,
                 betas=betas,
                 thinking=thinking,
+                window_dimension=window_dimension,
             )
         assistant_content = llm_response["content"]
         llm_caller.message_history.append({"role": "assistant", "content": assistant_content})
@@ -1454,8 +1458,8 @@ class ForgeAgent:
             task,
             step,
             assistant_content,
-            llm_caller.browser_window_dimension,
-            llm_caller.screenshot_resize_target_dimension,
+            window_dimension or llm_caller.browser_window_dimension,
+            llm_caller.get_screenshot_resize_target_dimension(window_dimension),
         )
         return actions
 

skyvern/forge/sdk/api/llm/api_handler_factory.py

@@ -491,6 +491,7 @@ class LLMCaller:
         tools: list | None = None,
         use_message_history: bool = False,
         raw_response: bool = False,
+        window_dimension: Resolution | None = None,
         **extra_parameters: Any,
     ) -> dict[str, Any]:
         start_time = time.perf_counter()
@@ -516,7 +517,21 @@ class LLMCaller:
             )
 
         if screenshots and self.screenshot_scaling_enabled:
-            screenshots = resize_screenshots(screenshots, self.screenshot_resize_target_dimension)
+            target_dimension = self.get_screenshot_resize_target_dimension(window_dimension)
+            if window_dimension and window_dimension != self.browser_window_dimension and tools:
+                # THIS situation only applies to Anthropic CUA
+                LOG.info(
+                    "Window dimension is different from the default browser window dimension when making LLM call",
+                    window_dimension=window_dimension,
+                    browser_window_dimension=self.browser_window_dimension,
+                )
+                # update the tools to use the new target dimension
+                for tool in tools:
+                    if "display_height_px" in tool:
+                        tool["display_height_px"] = target_dimension["height"]
+                    if "display_width_px" in tool:
+                        tool["display_width_px"] = target_dimension["width"]
+            screenshots = resize_screenshots(screenshots, target_dimension)
 
         await app.ARTIFACT_MANAGER.create_llm_artifact(
             data=prompt.encode("utf-8") if prompt else b"",
@@ -683,6 +698,11 @@ class LLMCaller:
 
         return parsed_response
 
+    def get_screenshot_resize_target_dimension(self, window_dimension: Resolution | None) -> Resolution:
+        if window_dimension and window_dimension != self.browser_window_dimension:
+            return get_resize_target_dimension(window_dimension)
+        return self.screenshot_resize_target_dimension
+
     async def _dispatch_llm_call(
         self,
         messages: list[dict[str, Any]],

skyvern/utils/image_resizer.py

@@ -22,7 +22,9 @@ def get_resize_target_dimension(
     ratio = window_size["width"] / window_size["height"]
     for dimension in max_scaling_targets.values():
         if abs(dimension["width"] / dimension["height"] - ratio) < 0.02:
-            return dimension
+            if dimension["width"] < window_size["width"]:
+                # we only return the dimension if it's smaller than the window size
+                return dimension
     return window_size
 
 

skyvern/webeye/scraper/scraper.py

@@ -15,6 +15,7 @@ from skyvern.constants import BUILDING_ELEMENT_TREE_TIMEOUT_MS, DEFAULT_MAX_TOKE
 from skyvern.exceptions import FailedToTakeScreenshot, ScrapingFailed, UnknownElementTreeFormat
 from skyvern.forge.sdk.api.crypto import calculate_sha256
 from skyvern.forge.sdk.core import skyvern_context
+from skyvern.utils.image_resizer import Resolution
 from skyvern.utils.token_counter import count_tokens
 from skyvern.webeye.browser_factory import BrowserState
 from skyvern.webeye.utils.page import SkyvernFrame
@@ -237,7 +238,7 @@ class ScrapedPage(BaseModel):
     url: str
     html: str
     extracted_text: str | None = None
-
+    window_dimension: dict[str, int] | None = None
     _browser_state: BrowserState = PrivateAttr()
     _clean_up_func: CleanupElementTreeFunc = PrivateAttr()
     _scrape_exclude: ScrapeExcludeFunc | None = PrivateAttr(default=None)
@@ -540,9 +541,11 @@ async def scrape_web_unsafe(
     text_content = await get_frame_text(page.main_frame)
 
     html = ""
+    window_dimension = None
     try:
         skyvern_frame = await SkyvernFrame.create_instance(frame=page)
         html = await skyvern_frame.get_content()
+        window_dimension = Resolution(width=page.viewport_size["width"], height=page.viewport_size["height"])
     except Exception:
         LOG.error(
             "Failed out to get HTML content",
@@ -563,6 +566,7 @@ async def scrape_web_unsafe(
         url=page.url,
         html=html,
         extracted_text=text_content,
+        window_dimension=window_dimension,
         _browser_state=browser_state,
         _clean_up_func=cleanup_element_tree,
         _scrape_exclude=scrape_exclude,